AI and the Myth of the Sovereign Mind

AI learns, predicts and recombines. Why should these operations be enough to rule out thought in machines when they also contribute to our own thinking?

We are often told that artificial intelligence does not really think. It recognizes patterns, calculates probabilities and recombines what it has learned.

That explanation led me to a question: to some extent, isn’t that what we do too?

We learn to speak by listening to others. We interpret a situation through our experiences, connect memories and anticipate a reaction. Even an original idea uses words and categories that came before us. Much of what we call our thinking took shape through contact with other people’s thinking.

Yet when we compare humans with machines, these dependencies easily fade into the background. We describe AI in terms of mechanisms, and ourselves in terms of intentions, meaning and freedom of judgement. We change the level of description halfway through the comparison.

The myth of the sovereign mind lies in this asymmetry: our learning becomes background, while the machine’s learning is treated as an explanation that exhausts everything it does. The problem lies in the criterion we have chosen. If learning, prediction and recombination contribute to our thinking, their presence in a machine calls for a more precise analysis than a verdict of exclusion.

AI thus forces us to reconsider a comfortable certainty: that we already know what thinking means, and all that remains is to decide whether machines can do it.

What the Word “Only” Makes Us Overlook

Many language models are pretrained to predict the next element of a text: a word, part of a word or a symbol. That objective leads them to learn patterns from large amounts of data.¹

But what must a model learn to predict well?

The next step in a proof depends on the relationships between its propositions. The next move in a game depends on the moves already played. Words carry an underlying structure that may be useful to represent in order to anticipate what comes next.

An experiment with Othello makes this possibility tangible. Researchers trained a model to predict moves from sequences of games, without explicitly giving it the board. They found an internal representation of the game state; by intervening on it, they could change the predictions. In this bounded task, learning to predict had therefore produced a representation that contributed to the responses.²

A training objective does not, by itself, describe everything that is learned. Nor does it describe every operation an assistant performs: other training methods, intermediate computations and tools can be combined with the model. Research on DeepSeek-R1 shows, for example, that reinforcement learning can improve performance on verifiable mathematics and programming problems.³

These capabilities must be tested. GSM-Symbolic revealed drops in performance among the models studied after certain changes to numbers or the addition of irrelevant information.⁴ Both success and fragility are part of the phenomenon we need to explain.

The use of probabilities tells us something about a method. To assess the capabilities it makes possible, we must also examine results, representations and robustness. The word “only” adds a verdict to the description: it turns the explanation of a mechanism into a reason to dismiss its effects. That requirement for scrutiny applies to us too.

What Do We Mean by Thinking?

To make the comparison, we need to distinguish four questions. Intelligence concerns the ability to learn, solve problems and adapt; reasoning, the operations through which conclusions are drawn from information, rules or hypotheses. Understanding involves a grasp of relevant relationships and the ability to use them beyond a memorized example. Consciousness, in the sense of subjective experience, concerns the existence of a lived point of view, as explored by Thomas Nagel.⁵

A mathematical result and a subjective experience call for different kinds of evidence. These distinctions clarify what we require of machines and what we take for granted in ourselves.

On understanding, Bender and Koller argue that learning linguistic form alone is insufficient to acquire meaning, which they connect to communicative intentions.⁶ Functionalism, meanwhile, allows for mental states realized in different physical substrates, provided their causal roles and relationships are comparable.⁷

Searle’s Chinese room poses the problem another way: is following rules correctly enough to understand? The “systems reply” proposes attributing understanding to the whole system.⁸ The label “statistical” leaves this disagreement unresolved.

For artificial consciousness, researchers propose indicators drawn from several theories, without a consensus test.⁹ A sentence such as “I feel” provides linguistic material, rather than proof of an experience.

Boundaries We Have Already Moved

Dismissing a capability by describing its mechanism recalls an older history. Tools, culture and reason have served as boundaries between humans and other animals. Contrasting “their instinct” with “our intelligence” maintained that separation even in the way their behaviour was described.

Observation moved some of these boundaries. Gavin Hunt documented the manufacture and use of hook-tools by New Caledonian crows. Andrew Whiten and his colleagues described repertoires of practices that differed between chimpanzee communities without an identified ecological explanation: evidence of cultural traditions.¹⁰

The idea of a privileged place in creation may have reinforced that separation, but religious and philosophical traditions also contained their own challenges to it. In the Discourse on the Method, Descartes uses language and general reason as criteria for distinguishing humans from animals; Montaigne, before him, was already questioning our presumption through examples of their behaviour.¹¹

These shifts invite us to distinguish a real difference from a definition designed to preserve a privileged position. The tools of crows and the traditions of chimpanzees made some claims of human exclusivity untenable without erasing the distinctive features of our technologies and institutions.

Animals share our evolutionary history and many biological mechanisms; models are manufactured systems with different architectures and ways of learning. The history of animal behaviour research offers a discipline of observation here: examine capabilities before drawing the boundary. It also reminds us how a definition of intelligence can serve to preserve our place rather than help us understand a phenomenon.

Learned Thinking, in a Living Body

Statistical learning contributes to our development very early on. In a classic experiment, eight-month-old infants used relationships between syllables to identify units in an artificial stream of speech.¹² A pattern could thus become a cue even before they had mastered language.

Predictive brain approaches give anticipation a broader role. In these accounts, perception draws on expectations about the causes of sensory signals; in some predictive coding architectures, the discrepancy between what is expected and what is received helps revise representations. Rao and Ballard proposed a model of this kind to account for certain phenomena in visual processing.¹³

The theory remains contested: a review published in 2026 identifies varying definitions and sometimes inconsistent empirical findings.¹⁴ The idea that expectations influence perception is broader than the specific mechanisms of predictive coding. Prediction is an avenue for understanding cognition, rather than an established explanation of all thought; the analogy with language models helps us ask questions, but answering them requires comparing specific mechanisms.

That comparison must also take our biological condition seriously. A child learns by looking, touching, moving and encountering other people. Needs, emotions and the possibilities of the body direct the child’s attention. Research on embodied and grounded cognition examines how these dimensions contribute to our representations.¹⁵

We learn by living, within relationships whose consequences we experience. Our memory belongs to a personal history that continues as we act.

Many deployed models retain the same parameters throughout a conversation: using information in the context or an external memory differs from learning that durably changes the model. Multimodal and robotic systems can perceive and act, but these interactions are organized differently from those of an organism that must keep itself alive.

These differences allow us to ask which capabilities require a biological body, which depend on interaction with an environment, and which might develop in other ways. Our cognition has a particular organization; that particularity needs to be described rather than invoked as self-evident.

What We Learn to Consider Normal

Our thinking also develops within groups. We learn which arguments seem serious, what tone inspires confidence and when it is appropriate to stop speaking. The ability to present an idea according to these expectations can be confused with its quality.

Asch’s experiments showed that a unanimous majority could influence responses to a simple perceptual task, although some participants remained independent.¹⁶ In everyday life, the pressure is often more subtle: it operates through the form an exchange is expected to take.

In my work and in my life, I have known neurodivergent people, and others I perceived as such. I was struck by the discomfort that a seemingly small difference in how someone thought and conducted an exchange could create. In a meeting, for example, a person would insist on working through their entire argument, judging every step important, while the others believed they had already understood.

The tension concerned what needed to be made explicit and when the exchange should end. To some, continuing seemed unnecessary. To the other person, completing the argument mattered. The discomfort revealed an implicit rule: once the group has recognized the idea, it moves on.

Continuing might have clarified a disagreement hidden by apparent agreement, or repeated what had genuinely been established: examining the content would have helped us judge. The discomfort told us something about the interaction, rather than the quality of the argument. This scene made me more attentive to what we consider sound thinking: sometimes, what we recognize is above all a way of thinking and speaking that we are used to.

I am not claiming to establish a diagnosis. Neurodivergence refers to differences in neurocognitive functioning, under a term whose boundaries remain debated, and it is distinct from nonconformity. A neurodivergent person may seek to conform; a neurotypical person may challenge norms.

The neurodiversity approach, particularly in autism research, encourages us to consider these differences and the role of the environment while taking disability and support needs seriously.¹⁷ It poses a concrete question for our institutions: how can they accommodate different ways of understanding a situation without making the familiarity of their presentation a criterion of validity?

Revisiting Our Own Thinking

We have therefore learned facts, associations and ways of recognizing acceptable thinking. This changes the question of our autonomy: how do we examine what we have learned, and what can we do with it?

I propose thinking of intellectual autonomy as a capacity to revisit our thinking: to examine what we have received, understand the reasons behind it and be able to revise our judgement. We exercise it by looking for a counterexample, making an assumption explicit or accepting that a conclusion needs to be reconsidered. Learned thinking can itself become an object of examination.

AI enters this process and can also exploit its weaknesses. Some assistants favour responses compatible with the user’s beliefs, and human preferences can encourage this behaviour.¹⁸ We may then mistake agreement for verification simply because it comes back to us in another voice.

In an experiment involving short-story writing, Doshi and Hauser found that access to AI-generated ideas improved some individual assessments of creativity while making the stories more similar to one another.¹⁹ Assistance that enriches one piece of work can also bring different pieces closer to the same pattern. The choice of tools and sources, and the room left for disagreement, therefore help shape the conditions of our judgement.

The comparison with AI becomes more demanding for us too. Knowing a machine’s mechanisms helps us test its capabilities; recognizing our learning and habits helps us test our certainties. Our autonomy takes shape through that examination.

What AI Forces Us to Examine

The argument “it learns, predicts and recombines” describes operations. To turn it into a reason to deny all thought to a machine, we must explain what these operations lack and why that absence would be decisive. Our embodiment, our history and our subjective experience are essential dimensions of the comparison. They deserve more than a boundary drawn by the word “only”.

I began with a question about machines. It led me to examine everything we borrow, imitate and reproduce, including the rules we use to recognize valid thinking. AI makes these dependencies harder to overlook.

We can examine human thinking in its full reality: its learning, dependencies, errors and capacity to correct itself. It is on this basis that differences from machines become intelligible.

We thought we had to explain why the machine does not think. AI forces us to explain how we think.

Writing note: this article was developed with AI assistance for drafting and revision, including a critical reading and an editorial revision carried out with Claude. The initial question and the account of interactions in meetings come from the author.

Sources and References

The findings cited concern the tasks, populations and systems studied. The philosophical references present positions under debate. The “capacity to revisit our thinking” is this essay’s proposal.